Ask a company for its process documentation and you'll get one of three things: nothing, a wiki page last edited two ERP versions ago, or a beautiful SOP that describes how the process was designed rather than how it's done. Ask the person who actually does the work, and you'll get the real thing, including the parts that were never written down anywhere.
That gap has a name in every back office: tribal knowledge. And it's why we made a deliberate bet at AltOps: the unit of teaching is a screen recording with narration, not a flowchart, not a form, not an integration spec.
What a recording captures that an SOP can't
Watch a skilled operator run a workflow with the mic on and you learn things no document contains:
- The real sequence, not the official one. The tab they always open first "just to check." The field the SOP doesn't mention because it predates it.
- The exceptions, in the wild. "This customer's POs always come in sizes-across-the-top, so I flip them." "If the price differs by under a dollar I take ours; over that, I email Dana."
- The judgment thresholds. When to proceed, when to flag, when to stop. This is the part every failed automation project skipped, and it's spoken, not written, knowledge.
- The current screens. A recording is documentation that's true as of the day it was made, of the actual UI, at the actual resolution, with the actual pop-ups.
This isn't a niche problem to route around. 92% of enterprises still depend on legacy systems for core operations, and the average enterprise runs hundreds of applications with barely a quarter of them integrated. The truth of most workflows lives on screens and in heads. The recording captures both at once.
Why now, technically
Teaching software by demonstration only works if the student can actually watch. Two years ago it couldn't: in April 2024, the best computer-use models completed about 12% of real desktop tasks on the OSWorld benchmark. By late 2025 agents had crossed the human baseline of ~72%, and the top models now sit in the mid-80s. Anthropic, OpenAI, and Google all shipped production computer-use models in that window; the capability went from demo to dependable.
What the benchmarks don't hand you is the job. A frontier model can operate a screen; it doesn't know that your team never ships partial quantities to that account, or which portal password lives where, or that the third Thursday is cutoff. The recording is how that knowledge gets in.
Capability came from the models. Competence comes from your people. The recording is the handoff between them.
From recording to running
The loop, end to end: an employee records one real pass of the workflow, talking through what they're doing and why. AltOps turns that into a skill (a structured understanding of the task, its systems, its rules and its stop points), and an agent runs it on its own cloud computer, in the same tools, while your team watches the live desktop. When it hits something the recording didn't cover, it asks instead of guessing. When someone corrects it, the correction becomes part of the skill.
That last part is the quiet payoff. The company's most fragile asset, the way work actually gets done, held in the heads of the people who do it, starts accumulating somewhere durable. People still leave, get promoted, go on holiday. The skill stays, current to the last correction, ready to teach the next agent or, for that matter, the next human.
SOPs were always a lossy copy of the real thing. It turns out the real thing was recordable all along.
Written by Kavish Soningra, teaching agents to run the back office at AltOps.
